A Competitive Array of Artificial Neural Networks for Use in Structural Impairment Detection

Abstract

In light of aging infrastructure, an opportunity exists to develop improved systems for monitoring and evaluation of infrastructure systems. Analytical modeling and experimental observation of an instrumented structure provide insight into to structural behavior. Such insight enables engineers and decision makers to maintain and monitor their structural systems. Accessible and practical dissemination of evaluations resulting from structural monitoring is a critical step in the successful management of infrastructure. As such, Structural Impairment Detection Systems (SIDS) target the detection of specific, expected impairments and provide succinct reports to decision makers. This paper details a methodology for use in Structural Impairment Detection (SID) that incorporates competitive arrays of artificial neural networks. Traditional modeling and instrumentation of a bridge structure are coupled with an advanced pattern recognition algorithm comprising arrays of competitive neural networks. Unknown variations in loading, material composition and interaction, and structural geometry make it difficult to model a physical structure exactly. While matching exact values for deflections or stresses between a physical structure and a computer model is not realistic outside of a simple laboratory experiment, one does expect that the overall behavior of a structure can be captured in a structural model, and trends in behavior can be validated. For this reason, neural networks are one appropriate approach to the problem of Structural Impairment Detections. Neural networks can be trained to examine trends in structural behavior and identify patterns that correspond to target impairments. The approach presented in this paper provides several competing, untrained networks the opportunity to be trained on individual data streams. The competition is carried out by implementing an orthogonality criterion that examines the output response of an individual network and the target response. This approach increases the speed of training and the ability of the neural networks to classify new data streams. A 94% correct classification rate on simulated data streams representing a complex physical draw bridge structure is achieved.